Stop When Further Reasoning Won't Help: Attention-State Adaptive Generation in Reasoning Models

In three linesASAG, a training-free method analyzing attention distributions, detects overthinking in reasoning models and adaptively stops generation. Tested on DeepSeek-R1-Distill and Qwen3, it improves accuracy by 3.2% while reducing generated tokens by 40% on Qwen3-8B.

Summary generated by Claude — human-verified